World Labs, the company Fei-Fei Li started after leaving Stanford, announced Atlas on September 1. They call it an omni world model for spatial intelligence, which sounds like marketing until you look at what it actually does. One model takes text, images, video, and 3D data as input, grounds all of it in a shared spatial context, and emits video, point clouds, and 3D Gaussian splats on the way out. If you have ever stitched together a video generator, a photogrammetry package, and a simulator for one project, that list should get your attention.
Camera geometry as a native input
The architectural choice that separates Atlas from the current crop of video generators is that camera poses are a first-class input type. Most video models interpret a text prompt and hope the pan or truck you described is what shows up in the render. Atlas takes exact camera geometry instead. You give it one to six reference images and a camera path, and it generates up to 60 seconds of 1440p video that follows that path. For production pipelines, this is the difference between prompting and programming. A repeatable camera input can be tested, versioned, and diffed. A prompt cannot.
Under the hood it is an autoregressive diffusion transformer trained with rectified flow, pretrained from scratch on all four modalities. The serving stack will look familiar to anyone who has run LLM inference at scale: KV caching, cache-aware routing, disaggregated serving, diffusion distillation. The team is clearly reusing the infrastructure playbook from language models, and the spatial context works a lot like a context window, with the same scaling limits.
Reconstruction from almost nothing
On the reconstruction side, Atlas claims faithful 3D reconstruction from as few as two or three input images, and it can hold more than a hundred posed images in its spatial context for larger scenes. Output includes explicit geometry: point clouds and 3D Gaussian splats, not just novel view renders. That explicit 3D output is what ties it into Marble, World Labs’ existing product, and future Marble versions will be built on Atlas.
The company also positions this for robotics. The Real-to-Sim workflow reconstructs a scene from roughly 24 frames of phone video, then generates RGB and depth sensor views from a simulated robot’s camera. If that holds up outside the demo videos, it removes a painful manual step in robot training pipelines. It is also the claim with the least public evidence so far, so treat it accordingly.
The benchmarks are self-published
World Labs reports human raters preferring Atlas’s camera-controlled generation over five competing models, with preference rates from 75% against MiniMax H3 up to 94% against Seedance 2.5. On 3D reconstruction, they report a mean absolute relative pointmap error of 25.3 versus 28.7 for Pi3X, the best open source specialist they compared against. Two caveats matter. Every number is company-run. And there is no technical report yet, so nobody outside World Labs has replicated any of it. The direction is plausible, the magnitudes are unverified.
Availability: a waitlist, not a product
Atlas is in early access with select partners via a request form. No pricing, no general availability date, no API documentation. The HN thread, where World Labs engineers were answering questions directly, spent a lot of time on where Atlas actually sits. The company’s own framing puts it between a renderer and a simulator: it predicts consistent 3D views and can output geometry, but it does not natively predict actions the way a simulator would. That is an honest and useful distinction.
What this means if you build with 3D or video
The interesting primitive here is spatial context as an interface. Placing a handful of reference images in 3D space and having the model interpolate between them is genuinely new, and consuming camera poses as input makes camera control deterministic in a way prompt-based video models are not. If your work touches virtual production, robotics simulation, or 3D content pipelines, the practical move is to prototype your evaluation against current tools now so you have a baseline when access opens. Then join the early access form and wait.
The competitive picture is crowded and the definitions are all over the place. Google’s Genie 3, Runway Solaris, Odyssey, Niantic Spatial, and AMI Labs all get called world models, and they do meaningfully different things. When everyone claims the same category with different capabilities, watch the data formats and ecosystem lock-in around things like Gaussian splats, not just the demo quality. Whichever formats become the interchange standard will matter more than which model tops the next benchmark.
My read: cautious optimism. The demos are strong, the architecture is sensible, and camera-native conditioning is the right bet for anyone who needs repeatable output. But this is a capability announcement with a waitlist attached. Do not build a dependency on it until there is a public API, real pricing, or a technical report you can read.